SAR despeckling by sparse reconstruction on affinity nets (SRAN)

This paper presents a new approach for multiplicative noise removal in SAR images based on sparse coding by dictionary learning and collaborative filtering. First, an affinity net is formed by clustering log-similar image patches where a cluster is represented as a node in the net. For each cluster, an under-complete dictionary is computed using the alternative decision method that iteratively updates the dictionary and the sparse coefficients. The nodes belonging to the same cluster are then reconstructed by a sparse combination of the corresponding dictionary atoms. The reconstructed patches are finally collaboratively aggregated to build the denoised image. Experimental results demonstrate superior despeckle filtering performance.